Strix (STRIX) — Agentic Framework & Deterministic vs. LLM-Driven Tooling

docs / design / agentic-framework

Strix — Agentic Framework & Deterministic vs. LLM-Driven Tooling

Technical audience, but written to be checkable by anyone: every claim below was verified with a direct search of the source (grep), not inferred from the project's marketing copy.

Strix does not use LangChain, LangGraph, CrewAI, or AutoGen

A repo-wide search (grep -rniE "langchain|langgraph|crewai|autogen" across every .py, .toml, and .md file) turns up exactly one match, and it is not a dependency: strix/skills/vulnerabilities/llm_prompt_injection.md — an internal knowledge-pack file that briefs Strix's own agents on how to test a target application for prompt-injection weaknesses, and mentions LangChain/LangGraph as examples of frameworks a target might be built with. It is content about what Strix can attack, not infrastructure Strix runs on.

Strix's actual dependency (pyproject.toml, [project] dependencies):

"openai-agents[litellm]>=0.19.0,<0.20",
"openai>=2.45.0,<3",
"litellm",

The OpenAI Agents SDK, with the LiteLLM extra enabled, is the entire agentic runtime. litellm (also listed directly) is what lets STRIX_LLM accept provider-prefixed model ids from OpenAI, Anthropic, Bedrock, Vertex, OpenRouter, and self-hosted endpoints, without Strix needing a separate adapter per provider. There is no separate graph-orchestration library (no LangGraph, no CrewAI) — Strix's own strix/core/agents.py (AgentCoordinator) is the orchestration layer, built directly on top of the SDK's single-agent primitives. See STRIX — Backends and HLD.md's "Third-party agent runtime" section for the full breakdown of what the SDK provides vs. what Strix built itself.

Why this matters for a reader coming from LangGraph/CrewAI

If you're familiar with LangGraph's explicit state-graph nodes/edges or CrewAI's Crew/Task/Process abstractions, the closest analogue in Strix is:

LangGraph/CrewAI concept Strix equivalent Where
Graph nodes / Crew agents SandboxAgent instances (root + children) strix/agents/factory.py
Graph edges / task handoffs Mailbox messages via AgentCoordinator.send() — not a static graph; the tree shape is decided at runtime by which agents call create_agent strix/core/agents.py, strix/tools/agents_graph/tools.py
Checkpointing / persistence Coordinator state snapshotted to agents.json; per-agent conversation state in a SQLite session (agents.db) strix/core/agents.py: _maybe_snapshot(), strix/core/sessions.py
Tool nodes Every entry in _BASE_TOOLS (strix/agents/factory.py) plus SDK-native Shell/Filesystem capabilities strix/tools/
A supervisor/router node The root agent, enforced by its system prompt (not a separate framework construct) strix/agents/prompts/system_prompt.jinja

The key structural difference: Strix's agent tree is not a predefined graph you author ahead of time (as in LangGraph). It's dynamically grown at runtime — the root agent decides how many specialists to spawn and what to name them, based on what it finds in the target, and the tree shape is only known after the scan finishes.

Deterministic vs. LLM-driven, module by module

This expands the four-question discipline used throughout docs/modules/*.md into one consolidated table, so "where is the AI, exactly" can be answered in one place.

Module Deterministic (no model call) LLM-driven
strix/interface Argument parsing, Docker preflight, run-directory naming, diff-scope computation, reading finished-run artifacts for the viewer warm_up_llm() performs a connectivity check only — no reasoning
strix/core Coordinator state snapshotting/resume, mailbox routing, budget/turn-limit enforcement (hooks.py), lifecycle-tool stop-condition checking None directly — run_agent_loop() drives the SDK's Runner, which is where the model actually reasons and calls tools
strix/agents Tool-list assembly, skill-markdown loading, Jinja template rendering None directly — this module prepares the prompt; the model reasoning happens once the agent runs
strix/tools CVSS metric validation/scoring math, output truncation/spilling, todo/notes CRUD persistence, mailbox message formatting, all SDK-native Shell/Filesystem execution web_search (Perplexity-backed) is the one tool that itself calls an external model API
strix/runtime 100% — container lifecycle, manifest/mount construction, resource limits, Caido guest-login retry loop None
strix/report CVSS score/severity arithmetic once metrics are supplied, vuln-NNNN ID assignment, Markdown/CSV/JSON/SARIF rendering, dependency/CVE-finding dedup (identity-based) report/dedupe.py: check_duplicate() — an explicit LLM-judge call for non-dependency findings, optionally on a distinct model (STRIX_DEDUPE_MODEL)
strix/config Settings precedence resolution, provider-quirk predicate functions None — this module decides how a model gets called, it doesn't call one
strix/skills 100% — static Markdown content, file lookup only None
Every actual testing decision (what to probe, which payload, when a finding is real) — Entirely the agent (root + children), via the SDK's Runner.run_streamed() turn loop

The pattern that holds everywhere: infrastructure, bookkeeping, and arithmetic are deterministic; judgment calls are LLM-driven, and the one deliberate exception (dedup) is called out explicitly rather than left ambiguous.

Related: STRIX — Backends, STRIX — Architecture (HLD), HLD.md, LLD.md.